Why Kimi K3 Is More Than Another Model Release

Kimi K3 landed this week as an open-weight model that early testers already rank as "Fable class," meaning it trades blows with the best closed systems. As we discussed earlier this week, the rule of thumb said open weights trail closed models by two years. That gap now looks closer to what we observed on the so called daily drivers (less powerful models that get the jobs done). But what does this shrinking gap mean for Anthropic's rumored IPO?

The gap between open and closed models was always a comforting lie

For some time, the industry repeated a soothing mantra: open weights are just toys. You were only taken seriously if you were a (western) closed lab shipping proprietary model.

Then DeepSeek happened. It demonstrated that frontier-quality research was no longer confined to a handful of Western labs. More importantly, it showed that open-weight releases could force the entire industry to recalibrate.

In the frontier space, however, there was a comfortable gap between closed and open models. Historically, it took roughly two years for an open weight model to reach the reasoning capabilities of the state of the art models.

When Anthropic released Fable, it appears that the game was over. Its hegemony as the world's most advanced AI lab was plain to all to see. They were at the top of the world weeks before their IPO.

Then, in a matter of weeks, GLM 5.2 was released. This open weight model was capable of trading blows with the best offerings from closed labs.

Colibri is a toy, and that is the point

GLM is free, but running it is not for the faint of heart. It requires datacenter-class GPUs. At half precision (FP8), its weights alone consume roughly 753 GB of RAM: a hardware class that is completely out of reach for everyday developers.

Now the part that looks unrelated but isn't. Project Colibri is a small open-source project, and I'll say plainly: it's a toy. It's not production infrastructure.

So why bring it up?

Colibri enables running GLM on a consumer-grade hardware.

By streaming the experts from disk, it allows running the behemoth of a model on a hardware that, according to its authors, cost less than the cooling system of datacenter GPUs.

Of course the performance is abysmal: it will output about 0.1 token per second.

But it represents a paradigm shift. It allows you to run the model that yesterday was impossible to run.

And the difference between 0.1 TPS and no token at all is greater than 0.1 TPS and 100 TPS.

What about Kimi K3

Kimi K3 shipped on web and app is another model from a Chinese lab (Moonshot) that seems to be able to face the very best closed offerings. Testers are calling K3 Fable class based on their own comparisons, which is a shorthand for a specific thing: it holds a coherent thread across long, messy, creative tasks the way the top closed models do. These tasks include:

• Generating fully playable, beautifully lit 3D Subway FPS zombie games;
• Creating flawless, 3D-printable V8 engine models;
• Building interactive browser operating systems with functional retro games.

You can check that out on this YouTube video showing Kimi in action:

To add to the offense, Kimi is bound to have its weight released 27th July.

So what? You might be wondering.

Kimi K3 proves that open offerings matching the best proprietary models within months of their release is now the norm, not the exception.

What this does to the Anthropic IPO story

Here's the uncomfortable part, and it's why any of this matters beyond the hobbyist threads.

An IPO is a bet on future cash flows. When investors price a frontier lab, they're implicitly pricing a promise. The pitch has always leaned on some version of "our models are meaningfully ahead, and staying ahead is expensive enough that nobody catches us."

Collapse the open-weight gap from two years to a few months, and that pitch gets harder to defend. Consider what changes:

• Pricing power erodes. If a free download is Fable class, the ceiling on what you can charge for general intelligence drops.
• The narrative shifts. The story has to move from "we have the best model" to "we have the best safety, tooling, reliability, and enterprise trust." Those are real and defensible. They're also a lower-margin, slower-growth story than raw capability leadership.
• The comparison set expands. Public-market analysts will benchmark a frontier lab against free alternatives that are visibly close. That's an awkward slide in any roadshow.

To be clear, none of this makes Anthropic a doomed IPO. Safety-first positioning, enterprise contracts, and genuine reliability advantages are worth real money.

But the valuation math built on a two-year capability promise needs revising. If open weights are months behind and closing, the defensible business is the boring, sticky, enterprise-services layer, not the model itself. Investors buying the 'we'll always be the bleeding edge' story are buying a melting asset. Investors buying the 'trusted enterprise AI platform' story might be buying something durable.

The two are priced very differently.

FAQ

What does "Fable class" actually mean?

It’s community shorthand, not an official benchmark. It means "competitive with the absolute best proprietary models on hard, long-context tasks."

Doesn't Anthropic still have an advantage in safety and enterprise reliability?

Yes, and that’s exactly the point. Anthropic’s actual advantage is no longer raw capability: Kimi K3 and GLM prove that is commoditizing. It is now trust, safety guardrails, and enterprise compliance. While these are highly defensible, they are not bleeding-edge tech monopoly margins.

Why does 0.1 tokens per second matter if it's practically unusable?

Because hardware scales faster than we think. Colibri running at 0.1 TPS on consumer hardware today proves the architecture works. If a weekend hacker can get it running at 0.1 TPS now, optimized versions will bring those gains to more capable gear. You don't judge a paradigm shift by its day-one speed; you judge it by the fact that the barrier to entry just hit zero.

Are you saying Anthropic's IPO will fail?

Not at all. Anthropic will likely have a highly successful IPO because enterprises desperately want a safe, reliable AI vendor. But if the market prices them like a company that holds a permanent lead in raw intelligence, that valuation is fragile. Pricing them as a highly trusted enterprise infrastructure company, it’s a much stronger.

Should my company drop closed models and switch to Kimi K3?

Don't switch yet (but architect so you can switch later).

The smartest move right now is to build an abstraction layer over your AI stack. Let your own latency, cost, and accuracy metrics make the decision, and ensure you aren't locked into a single vendor when the next open-weight drop happens in a few weeks.